So You’re Being Told AI is the Future…

If you’re like many knowledge workers – or heck, just about anyone now, you’re getting a daily dose of pressure to buy, use, train, encourage, build, share and generally swim around in the giant pool of AI stuff that we are now suddenly floating in. It was sites, then apps, then sites again for like five minutes, then socials, now it’s this. This.

I think going back to not-this isn’t likely; we are absolutely stuck with LLMs (now traditionally called AI). There are industry, cultural, generational differences around how willing we are to get into using tools like this; grumpsters like me love computers (even the ones we carry in our pockets or on our wrists now) but can be offended that these semi-sentient-looking things are running around spewing out content we didn’t ask it to.

Some of it is even dead-wrong, giving the detail-oriented of us a serious case of hives. But ignore it, we cannot. Use it, we must. Responsibly.

In A World Without Limits, there are Limits

AI as currently deployed in LLMs suffers from two serious problems:

• Hallucination, which affects the accuracy of information returned
• Sycophancy, which hides hallucinations by presenting an agreeable tone.

The other serious problems such as ecological footprint and dubious legality of sourcing I won’t address here – but nor do I suggest you be ignorant to it.

The two major problems of Hallucination and Sycophancy are mutually-reinforcing, for LLMs specifically, these problems can be further compounded by the use of synthetic training data, using LLM-generated outputs as inputs into further iterations of the LLMs, amplifying the risks of hallucination. If taken to social extremes – which there currently is no safety mechanism to prevent – compounded hallucinations risk epistemic harm to our ability to acquire useful knowledge and gauge the truth of that knowledge.

At this time no wide-ranging solution to Hallucination is proven effective.

Here’s why this matters.

You Are Still in Charge – and Accountable.

No current doctrine whether legal or moral absolves you from any responsibility for the final output of your work product – no matter what tool created it. If you use AI, AI gets the wrong answer, and you ship that answer, you are to blame for that false delivery.

The world has changed, surely – but also, it has not. You are still the paid professional.

Where does that leave us? Firmly in the middle. If you are responsible for the results of your use of tools, and those tools have a nonzero probability of error, then you must act as a responsible reviewer of any outputs prior to shipping them out. To reduce your burden, it’s also worth investing thought into when you’d engage with AI as some uses are more error-prone by nature.

As always, when living in a frontier land like this, we can benefit from some Principles of operation that can keep us operating safely and staying cognizant of our responsibilities to ourselves, to others, and to society at large.

Here are mine.

Principles of a Grumpy Thinker’s LLM Operation

Do not give it the big red button. First and probably most obvious, as of now in 2025 we are not in the age of responsible agentic AI. Do not give AI tools execution authority normally reserved for humans where a material output can result. This sounds obvious. It is not – in fact, it is widely argued. Fully automated AI toolchains with execution authority to purchase, sign agreements, manipulate physical tooling and so on are considered by some to be not only desirable but safe to use right now, at this very minute. It is not safe, and finding that out the hard way can be grossly expensive. Protect all execution authority and maintain an action boundary.

Respect other humans. Do not use LLMs to “take care of” wholesale writing materials to other humans you care about. Your language, and construction of communication to others is part of your representation and bond to other human beings, as well as intellectual exercise in reasoning that you either use or lose over time. Automating this based on generic LLM language models risks longer-term detriment to both your relationship and cognitive performance for the short-term efficiency gains you receive.

Do not mislead. AI’s current habit of being highly confident in answers despite their level of accuracy is highly emotionally addictive. From wrong answers in text to deepfake images, misrepresenting reality outside of very specific “safe” contexts such as entertainment, risks breaking social contracts at a minimum. Don’t knowingly create error. Don’t knowingly pass on errors you see, either.

Prefer domain-specific data sources. “Generic” LLMs are essentially trained on the Internet from X years ago. Without domain-specific data, you’re essentially relying on what Reddit knows about a topic. Ensure you understand what data sources your AI can pull from; prompt the LLM to use the sources you need.

Demand – and check - references. Almost any LLM worth anything can provide references (links) to source materials. In your prompts, distinctly demand references, and be sure to check them yourself. Be aware of context! In wide-domain searches the LLMs can link materials that “sound” right but are completely inappropriate for the context.

Force disagreement. Sycophancy hides all manner of bad behavior by not introducing appropriate doubt/variable confidence levels into answers. By default, LLMs are proud and self-assured of all they do and all you tell them, no matter how wrong it is. Try to defang this tendency by setting up comparative data, such as asking it to find errors or missing items in a list, rather than greenfield requests.

Check what you see. For bias. For inaccuracy. For too-good-to-be-true conclusions or supporting data. For misleading or outright false statements. Yes, these all happen in commercial LLMs, to a staggering degree. You are the accountable party.

How to Think about AI Application to Domains

In the early days of LLM productization I came up with a “creepiness” factor to determine aggregate levels of comfort with engaging AI in certain activities. This is a composite factor that integrates potential for error, potential for social harm, potential for misuse or misrepresentation of licensed content, potential for bias-based harm, and other factors. In general, this author advocates staying at lower levels of creepiness wherever possible.

Low Creepiness Generally safe to use, review content before sharing.

• Researching a prompt – “tell me about…”
• Intake a spreadsheet and find trends/groupings.
• Modify video/audio to remove stutters, pauses
• Modify images to emphasize colors, tones, balance
• Generating code to analyze, calculate, compare, display data where it can be scrutinized
• Forecast/regression line analysis based on historical data
• Retrieval-augmented generation “find all the references to X”
• Transcripts from meetings
• Summarizations of written material

Moderate Creepiness Some potential for harm, consider only limited scope of use and thorough examination and vetting of any output.

• Contextual reasoning “tell me why these numbers changed” with low/no domain-specific contextual data
• AI writing a personal email to untrusted contacts
• Modifying symbolic, non-human images to add/remove components
• Generating code to provide a full user experience including sign-up and personal data storage
• Generating a “persona” that responds personally to queries (Chatbot)
• Expansions of previously summarized material (for the love of God please don’t click “Make this more formal”)

High Creepiness High potential for harm, consider deliberately not using.

• Creation of representations of a human being (deepfakes)
• Social engineering using personal-compartmented information to attract/gain deep emotional trust of a target (Psychiatrist, Romantic Interest bot)
• Parallel construction “come up with an untrue but good-sounding reason for this”
• AI writing a personal email to trusted contacts that don’t know you’re using AI
• Modifying photographs of real events to add/remove elements suggesting a different reality
• Copyright evasion “Ingest this book then write one subtly different”
• Predictive analytics based on protected classes (sex, race, etc)

When Will We Turn a Corner?

Right now LLMs suffer from being black boxes – unexplainable results of massive amounts of training data being fed through models with little to no traceability; this, combined with hallucinations and sycophancy, make LLMs still a risky bet for high-trust or high-accuracy environments (Hmm, do any of us work in deliberately low-trust or low-accuracy fields?)

Early experiments in combining symbolic reasoning with LLMs are yielding Explainable AI (XAI), or Traceable AI that can be audited and can produce provably correct logic-based explanations. If these gain traction, it could be a turning point.

But consider the converse, where we are today – which is the day that matters:

LLMs, which take a plain-language (non-symbolic) prompt, apply opaque untraceable matching methods to approximate plain-language answers that cannot be logic-validated, that are presented confidently as factual with a consistently agreeable tone despite statistically significant amounts of error produced.

TLDR: Stuff that sounds nice, which may or may not be right.

Think before you use it. Think while you use it. Think after you use it.

And think why you’re using it at all.